把一次代理运行拆成 40 个决策，28 个交给 Jev
作者拆解单次代理运行：28 个决策走 Jev、9 个用确定性代码、仅 3 个需前沿模型，成本从 0.896 降到 0.0056 美元。
🚨 YOUR AGENT DOESN'T NEED A SMARTER MODEL it needs to stop waking one up 40 times per task I broke a single agent run into 40 decisions: → 28 can go through Jev → 9 belong in deterministic code → only 3 actually need a frontier LLM the insane part: those 28 decisions cost $0.8960 with a frontier model vs $0.0056 with Jev same job. ~160x less money burned on decisions that never needed generation in the first place. "is this evidence enough?" "which worker goes next?" "should we stop?" most agents are still renting GPT-6 / Opus-level reasoning for questions like these. that's the architecture shift nobody is talking about. I broke down the 10 decisions you're still overpaying a frontier model to make + exactly how to route them. article below ↓
这条是引用着另一条写的：x.com/0xkerazcity/status/2102042835352428590